This document summarizes and analyzes the interaction between intrinsic and extrinsic motivation from an economic and cognitive perspective. It discusses how performance incentives offered by an informed principal can impact an agent's perception of a task and their abilities. Incentives are shown to be only weak motivators in the short-run and potentially negative motivators in the long-run if they undermine intrinsic motivation. The document also examines how empowerment, help, and criticism impact motivation within relationships where parties have asymmetric information.

1. Review of Economic Studies (2003) 70, 489–520 0034-6527/03/00190489$02.00
c 2003 The Review of Economic Studies Limited
Intrinsic and Extrinsic Motivation
ROLAND B´ENABOU
Princeton University and Institute for Advanced Study
and
JEAN TIROLE
IDEI (Universit´e de Toulouse I), CERAS and MIT
First version received February 2000; final version accepted January 2003 (Eds.)
A central tenet of economics is that individuals respond to incentives. For psychologists and
sociologists, in contrast, rewards and punishments are often counterproductive, because they undermine
“intrinsic motivation”. We reconcile these two views, showing how performance incentives offered by an
informed principal (manager, teacher, parent) can adversely impact an agent’s (worker, child) perception
of the task, or of his own abilities. Incentives are then only weak reinforcers in the short run, and negative
reinforcers in the long run.We also study the effects of empowerment, help and excuses on motivation, as
well as situations of ego bashing reflecting a battle for dominance within a relationship.
Tom said to himself that it was not such a hollow world, after all. He had discovered a
great law of human action, without knowing it—namely, that in order to make a man or a
boy covet a thing, it is only necessary to make the thing difficult to attain. If he had been a
great and wise philosopher, like the writer of this book, he would now have comprehended
that Work consists of whatever a body is obliged to do, and that Play consists of whatever
a body is not obliged to do.
Mark Twain, The Adventures of Tom Sawyer (1876, Chapter 2).
INTRODUCTION
Should a child be rewarded for passing an exam, or paid to read a book? What impact do
empowerment and monitoring have on employees’ morale and productivity? Does receiving
help boost or hurt self-esteem? Why do incentives work well in some contexts, but appear
counterproductive in others? Why do people sometimes undermine the self-confidence of others
on whose effort and initiative they depend?
These questions will be studied here from a unifying perspective, emphasizing the interplay
between an individual’s personal motivation and his social environment. We shall thus model
the interactions between an agent with imperfect self-knowledge and an informed principal who
chooses an incentive structure, such as offering rewards and threatening punishments, delegating
a task, or simply giving encouragement, praise, or criticism.
It is a central theme of economics that incentives promote effort and performance, and there
is a lot of evidence that they often do (e.g. Gibbons (1997), Lazear (2000)). In other words,
contingent rewards serve as “positive reinforcers” for the desired behaviour. In psychology, their
effect is much more controversial. A long-standing paradigm clash has opposed proponents of
the economic view to the “dissonance theorists”, who argue that rewards may actually impair
performance, making them “negative reinforcers”, especially in the long run (see, e.g. Kruglanski
(1978) for an account of this debate, and Deci, Koestner and Ryan (1999) for a recent and
comprehensive meta-analysis of experimental results).
489

2. 490 REVIEW OF ECONOMIC STUDIES
Indeed, a substantial body of experimental and field evidence indicates that extrinsic moti-vation
(contingent rewards) can sometimes conflict with intrinsic motivation (the individual’s
desire to perform the task for its own sake). In a now classical experiment (see Deci, 1975),
college students were either paid or not paid to work for a certain time on an interesting puzzle.
Those in the no-reward condition played with the puzzle significantly more in a later unrewarded
“free-time” period than paid subjects, and also reported a greater interest in the task. This experi-ment
has since been replicated many times, with numerous variations in design (e.g.Wilson, Hull
and Johnson, 1981) and in types of subjects. For instance, similar effects were found for high-school
students in tasks involving verbal skills (Kruglanski, Friedman and Zeevi, 1971), and for
preschool children in activities involving drawing with new materials (Lepper, Greene and Nis-bett,
1973). In daily life, parents are quite familiar with what we shall call the “forbidden fruit”
effect: powerful or salient constraints employed by adults to enforce the prohibition of some
activity often decrease the child’s subsequent internalization of the adults’ disapproval.1 Kohn
(1993) surveys the results from a variety of programmes aimed at getting people to lose weight,
stop smoking, or wear seat belts, either offering or not offering rewards. Consistently, individ-uals
in “reward” treatments showed better compliance at the beginning, but worse compliance
in the long run than those in the “no-reward” or “untreated controls” groups. Taken together,
these many findings indicate a limited impact of rewards on “engagement” (current activity) and
a negative one on “re-engagement” (persistence).
A related body of work transposes these ideas from the educational setting to the workplace.
In well-known contributions, Etzioni (1971) argues that workers find control of their behaviour
via incentives “alienating” and “dehumanizing”, and Deci and Ryan (1985) devote a chapter of
their book to a criticism of the use of performance-contingent rewards in the work setting.2 And,
without condemning contingent compensation, Baron and Kreps (1999, p. 99) conclude that:
There is no doubt that the benefits of [piece-rate systems or pay-for-performance incentive
devices] can be considerably compromised when the systems undermine workers’ intrinsic
motivation.
Kreps (1997) reports his uneasiness when teaching human resources management and
discussing the impact of incentive devices in a way that is somewhat foreign to standard
economic theory. And indeed, recent experimental evidence on the use of performance-contingent
wages or fines confirms that explicit incentives sometimes result in worse compliance
than incomplete labour contracts (Fehr and Falk (1999), Fehr and Schmidt (2000), Gneezy
and Rustichini (2000a)). Relatedly, Gneezy and Rustichini (2000b) find that offering monetary
incentives to subjects for answering questions taken from an IQ test strictly decreases their
performance, unless the “piece rate” is raised to a high enough level. In the policy domain,
Frey and Oberholzer-Gee (1997) surveyed citizens in Swiss cantons where the government was
considering locating a nuclear waste repository; they found that the fraction supporting siting of
the facility in their community fell by half when public compensation was offered.
Our aim here will be twofold. First, we want to analyse the “hidden costs” of rewards
and punishments from an economic and cognitive perspective, rather than just posit an aversive
impact on motivation. Indeed, given that incentives work quite effectively in many instances, one
needs to understand in what cases they should be used with caution. More generally, we seek
to give a precise content to the loosely defined notions of intrinsic and extrinsic motivation, and
to clarify when, in the terminology of Frey (1997), the latter should be expected to “crowd out”
1. See, e.g. Lepper and Greene (1978). Relatedly, Akerlof and Dickens (1982) suggest that imposing stiffer
penalties for crimes might be counterproductive, if it undermines individuals’ “internal justification” for obeying the law.
2. See also Lepper and Greene (1978), Kohn (1993) and Frey (1997).

3. B´ENABOU & TIROLE INTRINSIC AND EXTRINSIC MOTIVATION 491
or “crowd in” the former. This information-based, strategic analysis distinguishes our approach
from Frey’s reduced-form treatment of these issues.
We consider an individual (the agent, “he”) who faces uncertainty about his payoff from
taking a particular action. The unknown variable could be a characteristic of the person himself,
such as raw ability, of the specific task at hand (long-run return, how difficult or enjoyable it
is to complete, etc.), or of the match between the two. Naturally, the agent will undertake the
task only if he has sufficient confidence in his own ability to succeed, and in the project’s net
return. As a result, people with a stake in his performance have strong incentives to manipulate
signals relevant to his self-knowledge. Given that effort and ability are usually complements
in the production of performance, they will want to boost his self-confidence, as well as his
interest in the task. Thus, in much of this paper, a principal (parent, spouse, friend, teacher, boss,
colleague, etc., “she”) has a vested interest in (derives a benefit from) the agent’s undertaking
and succeeding in the activity.
In many circumstances, both parties have private information about the agent’s suitability
to the task. The agent usually has better knowledge of his previous performances and of the
relevant circumstances (his effort intensity, the idiosyncratic factors that may have come into
play). He will often also receive privately signals about the attractiveness or unpleasantness of
the task, either from third parties (friends tell him that school is not fun, while cigarettes are
cool), from having performed similar ones in the past, or simply from his own experience as he
starts carrying out the current one. The principal, on the other hand, often has complementary
private information about the task or the agent’s prospects from it. For example, a teacher or
manager is better able to judge the difficulty of the subject or assignment, which, together with
the agent’s ability, conditions the probability of success. The principal may know better than the
agent whether the task is attractive, in terms of either being enjoyable to perform, or having a
high payoff for the agent. Last, while having less direct information about the agent’s previous
performances, she may be better trained at interpreting it due to her having performed the task
herself, or having seen many others attempt it. As we shall discuss later on, the observation that
others may have private information relevant to an individual’s self-view underlies several fields
of research in education and management. It is this type of private information that will be our
focus.3
In the first part of the paper we thus study the attributions made by an agent when a
principal with private information makes a decision, such as selecting a reward, delegating a
task or more simply encouraging the agent, that impacts the latter’s willingness to perform
the task. As was pointed out by Cooley (1902), the agent should then take the principal’s
perspective in order to learn about himself. The agent’s attribution of ulterior motivation to
the principal, or, in economics parlance, his attempt to infer her private information from
her decision, is what Cooley termed the “looking-glass self”. The influence of the principal’s
decision on the agent’s behaviour is then twofold: direct, through its impact on the agent’s payoff
from accomplishing the task (keeping information constant), and indirect, through his inference
process. In analysing intrinsic and extrinsic motivation we thus adopt a cognitive approach,
assuming that the individual seeks to extract from the words and deeds of those around him
signals about what they know that concerns him.4
3. Delfgauw and Dur (2002), in contrast, focus on the more standard case where workers have private information
about their own (dis)utility from working on the task, which they may then want to signal to, or conceal from, the
employer.
4. We in fact focus on the polar case where individuals are fully rational and Bayesian. Although people
surely make mistakes in processing information, we want the model to reflect the fact that they cannot constantly fool
themselves, or others.

4. 492 REVIEW OF ECONOMIC STUDIES
We first show that rewards may be only weak reinforcers in the short term and that, as
stressed by psychologists, they may have hidden costs, in that they become negative reinforcers
once they are withdrawn. The idea is that by offering low-powered incentives, the principal
signals that she trusts the agent. Conversely, rewards (extrinsic motivation) have a limited impact
on current performance, and reduce the agent’s motivation to undertake similar tasks in the future.
We then use the same logic to show that empowering the agent is likely to increase his intrinsic
motivation. Similarly, help offered by others may be detrimental to one’s self-esteem and create
a dependence.
More generally, we conclude that explicit incentives may, but need not, be negative
reinforcers; our analysis actually suggests when rewards and punishments work, and when they
backfire. The “crowding out” case first requires that the agent be less knowledgeable in some
dimension than the principal; this asymmetry of information is likely to be more important in
some settings (education, health, new occupations) than in others (relatively standardized jobs).
Furthermore, a sorting condition must hold, in that the principal must be more inclined to offer
a reward when the agent has limited ability or the task is unattractive. Otherwise, there will be
“crowding in”. Thus, when concerned about a potential negative impact of rewards, one should
first check whether the reward provider has private information about the task or the agent’s
talent. One should then, as the agent does, think through the provider’s ulterior motivation and
how her payoff from giving a contingent reward is affected by her knowledge.
In addition to low-powered incentives, we also investigate how the principal can sometimes
use non-contingent payments (similar to “burning money”) to signal her confidence in the agent’s
ability. While their short-term incentive effects differ, reducing the slope of the compensation
schedule (the piece rate) and increasing its base part (the fixed salary) are two related ways in
which the principal’s confidence-management motive will be reflected in equilibrium contracts.
Each has its domain of applicability (as we show), but both have similar effects on intrinsic, or
long run, motivation, as well as on wage inequality. Indeed, by weakening the link (elasticity)
between performance and compensation, both signalling strategies reduce earnings inequality
across workers, in a Lorenz sense.
While most of the social psychology and human resource management literatures emphasize
the necessity of boosting and protecting the self-esteem of one’s personal and professional
partners, people often criticize or downplay the achievements of their spouse, child, colleague,
coauthor, subordinate or teammate. In the second part of the paper we consider several reasons
why this may be, and formalize in more detail what is perhaps the most common one. We
argue that such “ego bashing” may reflect battles for dominance: by lowering the other’s self-confidence,
an individual may gain real authority within the relationship, enabling her to steer
joint decisions or projects in a preferred direction. This generally comes at a cost, however,
namely the risk of demotivating the partner from seeking good projects, or from exerting effort
at the implementation stage. We study this tradeoff, distinguishing two related forms of ego
bashing: one is “by omission”, where the principal omits to report news favourable to the agent;
the other is active “disparaging”, in which she explicitly belittles the agent. While both strategies
lower the agent’s self-confidence, the first one is reversible (the news can always be revealed later
on), whereas the second is not. This is shown to have interesting implications for the timing of
strategic disclosures of information (ego bashing and ego boosting) in situations where both the
agent’s initiative and joint control rights are at stake.
The paper is organized as follows. Section 1 provides a general introduction to the “looking-glass
self” mechanism. Section 2 analyses the interplay of intrinsic and extrinsic motivation,
focusing in particular on the hidden cost of rewards. Section 3 shows how the main insights
carry over to other confidence-management strategies such as delegation, help and coaching.
Section 4 studies the costs and benefits of ego bashing. Section 5 offers concluding remarks.

5. B´ENABOU & TIROLE INTRINSIC AND EXTRINSIC MOTIVATION 493
1. THE LOOKING-GLASS SELF
We begin here with a relatively general and abstract framework, then specialize it in the rest of
the paper. Readers more interested in specific psychological and economic applications than in a
unified presentation of the underlying mechanisms may want to proceed directly to Section 2.
There are two players, an agent (he) and a principal (she). The agent selects a continuous
action or effort level e that impacts both his and the principal’s utilities. The principal knows a
parameter

6. , such as the difficulty of the task or the agent’s ability to perform it, that affects the
agent’s payoffs from e. Thus informed, she selects a policy p (belonging to R, for expositional
simplicity) prior to the agent’s choice of action; this may be a wage or contingent reward, help,
surveillance, delegation, disclosure of information, or any other “extrinsic motivator” that can
affect, directly or indirectly, the agent’s behaviour. The agent’s and the principal’s payoffs are
denoted UA(

7. , e, p) and UP(

8. , e, p). Prior to his decision, the agent may also privately receive
a signal that is informative about

9. .We shall assume for simplicity that this signal is redundant
if one already knows the principal’s information (

10. is a sufficient statistic for (

11. , )), but none
of our main conclusions hinge on this assumption. What really matters is that the principal has
information relevant to the agent’s perception of himself or his task, and (for the specific “trust
effect” discussed below) that the principal be uncertain about the agent’s motivation. The timing
of the game is as follows:
Stage 1: The principal learns the parameter

12. and selects a policy p.
Stage 2: After observing the policy chosen by the principal and learning , the agent
chooses an action e.
Let us assume here (for notational simplicity) that the agent’s optimal action e depends only
on p and on his conditional expectation

13. ˆ(, p) of the unknown parameter.5 The conditioning
of

14. ˆ on p is the “looking-glass-self” phenomenon, whereby the agent tries to see through the
principal’s ulterior motives that led to p being selected. As long as the agent’s participation in
the relationship is not at stake, the principal’s expected payoff from choosing a policy p when
she has information

15. is thus
E [UP(

16. , e(p,

17. ˆ(, p)), p) |

18. ].
Assuming differentiability (again for simplicity), the principal’s choice of policy takes three
effects into consideration:
E
@UP
@p
+
@UP
@e
·
@e
@p
+
@UP
@e
·
@e
@

19. ˆ
·
@

20. ˆ
@p

25. = 0. (1)
The first term on the L.H.S. of (1) is the direct effect of p on the principal’s payoff. For example,
if the policy is a wage or bonus, as in the next section, this term is the direct cost of this
compensation, keeping the agent’s behaviour constant. The second term corresponds to the direct
impact of p on the agent’s behaviour. Thus, ceteris paribus, a bonus increases the incentive to
exert effort. These two effects have been investigated in detail in the agency literature.
We shall be interested in the third, more novel effect, which corresponds to the principal’s
confidence-management motive. Whenever the principal’s choice of policy is guided by private
information, the agent will update his beliefs in reaction to the choice of p (term @

26. ˆ/@p). The
principal must then take into account how the agent’s interpretation of her choice will affect
his self-confidence—that is, his perceived prospects from undertaking the task. A particularly
important issue is whether a higher level of self-confidence affects the agent’s decision making in
5. More generally, it will depend on p and on the conditional distribution of

27. , given (, p).

28. 494 REVIEW OF ECONOMIC STUDIES
a direction that the principal likes ((@UP/@e)(@e/@

29. ˆ) 0) or dislikes ((@UP/@e)(@e/@

30. ˆ)
0). Sections 2 and 3 will examine many common situations where the principal gains from
boosting the agent’s self-confidence. Section 4, on the other hand, will focus on cases where
she may be reluctant to enhance the agent’s self-confidence, or may even want to undermine it.
The confidence-management motive itself can itself arise through two channels, which we
term the profitability effect and the trust effect. The former arises when the agent’s type, on
which the principal has private information, enters the principal’s objective function in a way
that would lead her to offer different policies to different agents, even if it did not affect anyone’s
effort level. This differential profitability of a given policy across types corresponds to a standard
sorting condition; thus, for a one-dimensional policy it means that6
@
@

31. @UP/@p
@UP/@e
has a constant sign. (2)
Suppose, for instance, that an employer’s expected profits are (proportionally) more sensitive
to the employee’s ability when the latter is empowered to make decisions than when he is
closely monitored. The principal will then, ceteris paribus, delegate more to employees she
thinks more highly of, and delegation will be seen as good news by employees. In contrast,
no such profitability effect exists when the principal’s private knowledge concerns the cost of
accomplishing the task, or other aspects of it that bear solely on the agent’s utility, and not on her
own payoff.
The trust (or distrust) effect, on the contrary, arises when the principal’s private information
concerns a parameter, such as the cost or pleasure of accomplishing the task, that directly enters
only in the agent’s incentive problem—as envisioned by the principal. The issue here is how
confident the principal is as to the agent’s intrinsic motivation—that is, how she thinks the
agent perceives the task and his suitability to it. A principal who has bad news about the agent’s
parameter

32. will be pessimistic about the agent’s own signal , and may consequently fear that
he will not be motivated enough to exert effort in the absence of added incentives. Providing
stronger incentives, however, will at least partially reveal the principal’s damaging information
(compounding the signal ). Thus, once again, extrinsic motivation may “crowd out” intrinsic
motivation, and the optimal contract will be shaped by this tradeoff.
It is worth noting also that for the pure trust effect to operate, there must be some uncertainty
( ) on the part of the principal about the exact incentives perceived by the agent. Otherwise,
the latter’s response